Bayesian Change Point Detection in Monitoring Clinical Outcomes
A control chart is typically used to monitor the behaviour of a process over time, and to signal if a significant change in either the stability or dispersion of the process is detected. The signal can then be investigated to identify potential causes of the change and implement corrective or preventive actions (Montgomery 2008). This approach is well established in industry, and has more recently been taken up in other fields such as health care. In the clinical context, control charts can be used to monitor processes such as mortality after surgery. Here, patients comprise the units being monitored, and the underlying process parameter of interest is the mortality rate. One of the difficulties with this approach in a health-care context is that the process parameter is not constant: a patient's probability of death depends on their personal risk profile. In light of this, if is often preferable to construct a risk-adjusted control chart, that is one that takes into account patient mix (Cook et al. 2008; Grigg and Farewell 2004). For example, Steiner and Cook (2000) developed a risk-adjusted version of the cumulative sum control chart (CUSUM) to monitor surgical outcomes, death and survival, accounting for ...